Optimization of groundwater monitoring network using a combined data mining approach and NSGA-II genetic algorithm (Shamil-Ashkara aquifer, Hormozgan province)

Groundwater resources are essential for fulfilling the various needs of human societies and maintaining ecological equilibrium. They represent a substantial portion of the world’s accessible freshwater reserves and are widely used for agriculture, industry, and domestic purposes. However, the increasing global demand for water, coupled with the adverse impacts of climate change and pollution, has resulted in the unsustainable exploitation of these vital resources. Over-extraction of groundwater has led to severe consequences, including declining water tables, deteriorating water quality, well depletion, and land subsidence. These challenges necessitate the implementation of effective management strategies to ensure the sustainable utilization of groundwater resources. Accurate assessment and prediction of groundwater levels (GWL) are crucial for informed decision-making and sustainable groundwater management. Traditional physics-based models, while effective, often require extensive input data and precise hydrogeological parameters. Machine learning (ML) techniques offer an alternative approach, enabling GWL prediction with potentially limited data requirements. This study focuses on optimizing the existing piezometer network in the Shemil-Ashkara plain, located in northeastern Hormozgan province, Iran. The primary objective is to identify optimal piezometer locations and potentially eliminate redundant wells using multi-criteria decision-making (AHP) and principal component analysis (PCA). This approach leverages data mining and artificial intelligence techniques to enhance the efficiency and effectiveness of groundwater monitoring networks.
Materials and Methods:
The study area encompasses the Shemil-Ashkara plain, situated in northeastern Hormozgan province, Iran. The region is characterized by diverse geological formations, primarily Quaternary alluvial deposits, and a semi-arid climate with limited rainfall and high evaporation rates. The primary land uses in the area include rangeland, agriculture, and residential areas. The research methodology involved a multi-stage approach. Initially, the Analytic Hierarchy Process (AHP) method was employed to identify suitable locations for new observation wells. Eight criteria were selected based on a literature review, guidelines from the Ministry of Energy, and expert opinions: long-term average groundwater level, annual groundwater decline rate, slope of decline changes, density of exploitation wells, distance from the river, geological formation, land use, and distance from faults. Pairwise comparisons were conducted using a scale of 1 to 9 to determine the relative weights of each criterion. The consistency ratio (CR) was calculated to ensure the reliability of the expert judgments. Spatial analysis was performed using ArcGIS software to combine the weighted criteria maps and generate a final prioritization map for piezometer locations.
The existing monitoring network was evaluated by constructing Thiessen polygons for each well to delineate their respective areas of influence. The average location score for each polygon was calculated, and wells with scores below 0.3 were deemed unsuitable. To validate the selection of unsuitable wells, homogeneity tests were conducted on their hydrographic data using the Petits and Standard Normal Homogeneity Test (SNHT) methods. Principal component analysis (PCA) was applied to determine the relative importance of the existing wells. A sequential elimination approach was employed, where each well was removed iteratively, and the correlation between the remaining wells and the first principal component (PC1) was calculated. The average of these correlation coefficients represented the relative importance of each well. The uncertainty associated with removing less important wells was assessed by calculating the average coefficient of variation (CV) of annual groundwater levels. Kriging was used to interpolate groundwater levels under two scenarios: using all observation wells and removing less important wells identified by AHP and PCA. The Gaussian variogram model provided the best fit to the experimental data. The root mean squared error (RMSE) was calculated to evaluate the accuracy of the interpolation under both scenarios. Finally, the Non-dominated Sorting Genetic Algorithm II (NSGA-II) was employed to determine the optimal piezometer network configuration, considering the trade-off between the number of piezometers and interpolation accuracy. The algorithm used a binary chromosome representation, where each gene corresponded to the presence or absence of a well. The objective functions were to minimize the number of piezometers and the RMSE between observed and estimated groundwater levels. The Inverse Distance Weighting (IDW) method was used to estimate groundwater levels for each chromosome.
Results and Discussion:
The AHP analysis revealed that the southern and southeastern regions of the Shemil-Ashkara plain exhibited the highest priority for piezometer placement due to the concentration of agricultural activities, a high density of exploitation wells, a significant groundwater level decline, and high decline rates. The Thiessen polygon analysis identified five wells (W7, W9, W12, W14, and W16) as unsuitable due to their low location scores and inadequate coverage of the aquifer. Homogeneity tests confirmed the unsuitability of these wells, indicating non-homogeneous hydrographic data potentially caused by over-extraction of groundwater. PCA identified different sets of less important wells compared to AHP, reflecting the distinct methodologies and input data used by each technique. Removing less important wells based on AHP resulted in a decrease in the standard error of interpolation, while removing wells based on PCA increased the error. However, both methods led to an increase in RMSE, suggesting that removing these wells did not improve overall interpolation accuracy. The NSGA-II algorithm generated a Pareto front representing the trade-off between the number of piezometers and RMSE. The optimal network configuration, balancing cost and accuracy, consisted of 16 piezometers.
Conclusion and Suggestions:
This study demonstrated the effectiveness of the AHP method in identifying optimal locations for new piezometers in the Shemil-Ashkara plain. The integration of geological, hydrological, and land use factors enhanced the optimization process. While PCA provided insights into the relative importance of existing wells, removing less important wells based on either AHP or PCA did not improve the overall accuracy of groundwater level interpolation. The NSGA-II algorithm proved valuable in determining the optimal network configuration, considering both cost and accuracy objectives. It is recommended to apply the AHP method in other plains for optimizing groundwater monitoring networks. Further research should explore the inclusion of additional variables, such as ground-based measurements and outputs from groundwater simulation models, to refine the optimization process.